AMDGPUTargetLowering(TM) {
addRegisterClass(MVT::i1, &AMDGPU::SReg_64RegClass);
- addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
+ addRegisterClass(MVT::i64, &AMDGPU::VSrc_64RegClass);
- addRegisterClass(MVT::v2i1, &AMDGPU::VReg_64RegClass);
- addRegisterClass(MVT::v4i1, &AMDGPU::VReg_128RegClass);
-
- addRegisterClass(MVT::v16i8, &AMDGPU::SReg_128RegClass);
addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
- addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
- addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
-
- addRegisterClass(MVT::v1i32, &AMDGPU::VReg_32RegClass);
+ addRegisterClass(MVT::i32, &AMDGPU::VSrc_32RegClass);
+ addRegisterClass(MVT::f32, &AMDGPU::VSrc_32RegClass);
- addRegisterClass(MVT::v2i32, &AMDGPU::VReg_64RegClass);
- addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
- addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
+ addRegisterClass(MVT::f64, &AMDGPU::VSrc_64RegClass);
+ addRegisterClass(MVT::v2i32, &AMDGPU::VSrc_64RegClass);
+ addRegisterClass(MVT::v2f32, &AMDGPU::VSrc_64RegClass);
addRegisterClass(MVT::v4i32, &AMDGPU::VReg_128RegClass);
addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
setOperationAction(ISD::ADD, MVT::i64, Legal);
setOperationAction(ISD::ADD, MVT::i32, Legal);
+ setOperationAction(ISD::BITCAST, MVT::i128, Legal);
+
+ // We need to custom lower vector stores from local memory
+ setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
+ setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
+
setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
+ setOperationAction(ISD::ZERO_EXTEND, MVT::i64, Custom);
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
+ setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::f32, Custom);
+ setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v16i8, Custom);
+ setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::v4f32, Custom);
setLoadExtAction(ISD::SEXTLOAD, MVT::i32, Expand);
- setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
+ setTruncStoreAction(MVT::f64, MVT::f32, Expand);
+
+ setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
setTargetDAGCombine(ISD::SELECT_CC);
return VT.bitsGT(MVT::i32);
}
+bool SITargetLowering::shouldSplitVectorElementType(EVT VT) const {
+ return VT.bitsLE(MVT::i8);
+}
SDValue SITargetLowering::LowerParameter(SelectionDAG &DAG, EVT VT,
SDLoc DL, SDValue Chain,
MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
SDValue Ptr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
DAG.getConstant(Offset, MVT::i64));
- return DAG.getExtLoad(ISD::ZEXTLOAD, DL, VT, Chain, Ptr,
+ return DAG.getLoad(VT, DL, Chain, Ptr,
MachinePointerInfo(UndefValue::get(PtrTy)),
- VT, false, false, ArgVT.getSizeInBits() >> 3);
+ false, false, false, ArgVT.getSizeInBits() >> 3);
}
return MVT::i32;
}
+bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
+ VT = VT.getScalarType();
+
+ if (!VT.isSimple())
+ return false;
+
+ switch (VT.getSimpleVT().SimpleTy) {
+ case MVT::f32:
+ return false; /* There is V_MAD_F32 for f32 */
+ case MVT::f64:
+ return true;
+ default:
+ break;
+ }
+
+ return false;
+}
+
//===----------------------------------------------------------------------===//
// Custom DAG Lowering Operations
//===----------------------------------------------------------------------===//
switch (Op.getOpcode()) {
default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
case ISD::BRCOND: return LowerBRCOND(Op, DAG);
+ case ISD::LOAD: {
+ LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
+ if (Load->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS &&
+ Op.getValueType().isVector()) {
+ SDValue MergedValues[2] = {
+ SplitVectorLoad(Op, DAG),
+ Load->getChain()
+ };
+ return DAG.getMergeValues(MergedValues, 2, SDLoc(Op));
+ } else {
+ return SDValue();
+ }
+ }
case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, DAG);
+ case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, DAG);
case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
case ISD::INTRINSIC_WO_CHAIN: {
unsigned IntrinsicID =
case Intrinsic::r600_read_tidig_z:
return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
AMDGPU::VGPR2, VT);
-
+ case AMDGPUIntrinsic::SI_load_const: {
+ SDValue Ops [] = {
+ ResourceDescriptorToi128(Op.getOperand(1), DAG),
+ Op.getOperand(2)
+ };
+
+ MachineMemOperand *MMO = MF.getMachineMemOperand(
+ MachinePointerInfo(),
+ MachineMemOperand::MOLoad | MachineMemOperand::MOInvariant,
+ VT.getSizeInBits() / 8, 4);
+ return DAG.getMemIntrinsicNode(AMDGPUISD::LOAD_CONSTANT, DL,
+ Op->getVTList(), Ops, 2, VT, MMO);
+ }
+ case AMDGPUIntrinsic::SI_sample:
+ return LowerSampleIntrinsic(AMDGPUISD::SAMPLE, Op, DAG);
+ case AMDGPUIntrinsic::SI_sampleb:
+ return LowerSampleIntrinsic(AMDGPUISD::SAMPLEB, Op, DAG);
+ case AMDGPUIntrinsic::SI_sampled:
+ return LowerSampleIntrinsic(AMDGPUISD::SAMPLED, Op, DAG);
+ case AMDGPUIntrinsic::SI_samplel:
+ return LowerSampleIntrinsic(AMDGPUISD::SAMPLEL, Op, DAG);
+ case AMDGPUIntrinsic::SI_vs_load_input:
+ return DAG.getNode(AMDGPUISD::LOAD_INPUT, DL, VT,
+ ResourceDescriptorToi128(Op.getOperand(1), DAG),
+ Op.getOperand(2),
+ Op.getOperand(3));
}
}
}
return Chain;
}
+SDValue SITargetLowering::ResourceDescriptorToi128(SDValue Op,
+ SelectionDAG &DAG) const {
+
+ if (Op.getValueType() == MVT::i128) {
+ return Op;
+ }
+
+ assert(Op.getOpcode() == ISD::UNDEF);
+
+ return DAG.getNode(ISD::BUILD_PAIR, SDLoc(Op), MVT::i128,
+ DAG.getConstant(0, MVT::i64),
+ DAG.getConstant(0, MVT::i64));
+}
+
+SDValue SITargetLowering::LowerSampleIntrinsic(unsigned Opcode,
+ const SDValue &Op,
+ SelectionDAG &DAG) const {
+ return DAG.getNode(Opcode, SDLoc(Op), Op.getValueType(), Op.getOperand(1),
+ Op.getOperand(2),
+ ResourceDescriptorToi128(Op.getOperand(3), DAG),
+ Op.getOperand(4));
+}
+
SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
return DAG.getNode(ISD::BUILD_PAIR, DL, VT, Op.getOperand(0), Hi);
}
+SDValue SITargetLowering::LowerZERO_EXTEND(SDValue Op,
+ SelectionDAG &DAG) const {
+ EVT VT = Op.getValueType();
+ SDLoc DL(Op);
+
+ if (VT != MVT::i64) {
+ return SDValue();
+ }
+
+ return DAG.getNode(ISD::BUILD_PAIR, DL, VT, Op.getOperand(0),
+ DAG.getConstant(0, MVT::i32));
+}
+
//===----------------------------------------------------------------------===//
// Custom DAG optimizations
//===----------------------------------------------------------------------===//
return false;
}
+const TargetRegisterClass *SITargetLowering::getRegClassForNode(
+ SelectionDAG &DAG, const SDValue &Op) const {
+ const SIInstrInfo *TII =
+ static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
+ const SIRegisterInfo &TRI = TII->getRegisterInfo();
+
+ if (!Op->isMachineOpcode()) {
+ switch(Op->getOpcode()) {
+ case ISD::CopyFromReg: {
+ MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
+ unsigned Reg = cast<RegisterSDNode>(Op->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(Reg)) {
+ return MRI.getRegClass(Reg);
+ }
+ return TRI.getPhysRegClass(Reg);
+ }
+ default: return NULL;
+ }
+ }
+ const MCInstrDesc &Desc = TII->get(Op->getMachineOpcode());
+ int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
+ if (OpClassID != -1) {
+ return TRI.getRegClass(OpClassID);
+ }
+ switch(Op.getMachineOpcode()) {
+ case AMDGPU::COPY_TO_REGCLASS:
+ // Operand 1 is the register class id for COPY_TO_REGCLASS instructions.
+ OpClassID = cast<ConstantSDNode>(Op->getOperand(1))->getZExtValue();
+
+ // If the COPY_TO_REGCLASS instruction is copying to a VSrc register
+ // class, then the register class for the value could be either a
+ // VReg or and SReg. In order to get a more accurate
+ if (OpClassID == AMDGPU::VSrc_32RegClassID ||
+ OpClassID == AMDGPU::VSrc_64RegClassID) {
+ return getRegClassForNode(DAG, Op.getOperand(0));
+ }
+ return TRI.getRegClass(OpClassID);
+ case AMDGPU::EXTRACT_SUBREG: {
+ int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
+ const TargetRegisterClass *SuperClass =
+ getRegClassForNode(DAG, Op.getOperand(0));
+ return TRI.getSubClassWithSubReg(SuperClass, SubIdx);
+ }
+ case AMDGPU::REG_SEQUENCE:
+ // Operand 0 is the register class id for REG_SEQUENCE instructions.
+ return TRI.getRegClass(
+ cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue());
+ default:
+ return getRegClassFor(Op.getSimpleValueType());
+ }
+}
+
/// \brief Does "Op" fit into register class "RegClass" ?
bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
unsigned RegClass) const {
-
- MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
- SDNode *Node = Op.getNode();
-
- const TargetRegisterClass *OpClass;
const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
- if (MachineSDNode *MN = dyn_cast<MachineSDNode>(Node)) {
- const SIInstrInfo *TII =
- static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
- const MCInstrDesc &Desc = TII->get(MN->getMachineOpcode());
- int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
- if (OpClassID == -1) {
- switch (MN->getMachineOpcode()) {
- case AMDGPU::REG_SEQUENCE:
- // Operand 0 is the register class id for REG_SEQUENCE instructions.
- OpClass = TRI->getRegClass(
- cast<ConstantSDNode>(MN->getOperand(0))->getZExtValue());
- break;
- default:
- OpClass = getRegClassFor(Op.getSimpleValueType());
- break;
- }
- } else {
- OpClass = TRI->getRegClass(OpClassID);
- }
-
- } else if (Node->getOpcode() == ISD::CopyFromReg) {
- RegisterSDNode *Reg = cast<RegisterSDNode>(Node->getOperand(1).getNode());
- OpClass = MRI.getRegClass(Reg->getReg());
-
- } else
+ const TargetRegisterClass *RC = getRegClassForNode(DAG, Op);
+ if (!RC) {
return false;
-
- return TRI->getRegClass(RegClass)->hasSubClassEq(OpClass);
+ }
+ return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
}
/// \brief Make sure that we don't exeed the number of allowed scalars
/// \brief Fold the instructions after slecting them
SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
SelectionDAG &DAG) const {
+ const SIInstrInfo *TII =
+ static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
Node = AdjustRegClass(Node, DAG);
- if (AMDGPU::isMIMG(Node->getMachineOpcode()) != -1)
+ if (TII->isMIMG(Node->getMachineOpcode()))
adjustWritemask(Node, DAG);
return foldOperands(Node, DAG);
/// bits set in the writemask
void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
SDNode *Node) const {
- if (AMDGPU::isMIMG(MI->getOpcode()) == -1)
+ const SIInstrInfo *TII =
+ static_cast<const SIInstrInfo*>(getTargetMachine().getInstrInfo());
+ if (!TII->isMIMG(MI->getOpcode()))
return;
unsigned VReg = MI->getOperand(0).getReg();
switch (N->getMachineOpcode()) {
default: return N;
- case AMDGPU::REG_SEQUENCE: {
- // MVT::i128 only use SGPRs, so i128 REG_SEQUENCEs don't need to be
- // rewritten.
- if (N->getValueType(0) == MVT::i128) {
- return N;
- }
- const SDValue Ops[] = {
- DAG.getTargetConstant(AMDGPU::VReg_64RegClassID, MVT::i32),
- N->getOperand(1) , N->getOperand(2),
- N->getOperand(3), N->getOperand(4)
- };
- return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::i64, Ops);
- }
-
case AMDGPU::S_LOAD_DWORD_IMM:
NewOpcode = AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
// Fall-through